1. science
2. mathematics
3. statistics

YASS: Yet Another Steganographic Scheme
that Resists Blind Steganalysis
Kaushal Solanki†† , Anindya Sarkar† , and B. S. Manjunath†
†Department of Electrical and Computer Engineering,⋆⋆
University of California,
Santa Barbara, CA 93106
††Mayachitra Inc.,
5266 Hollister Avenue,
Santa Barbara, CA 93111
solanki@mayachitra.com,anindya@ece.ucsb.edu,manj@ece.ucsb.edu
Abstract. A new, simple, approach for active steganography is proposed in this paper that can successfully resist recent blind steganalysis methods, in addition to surviving distortion constrained attacks. We
present Yet Another Steganographic Scheme (YASS), a method based
on embedding data in randomized locations so as to disable the selfcalibration process (such as, by cropping a few pixel rows and/or columns
to estimate the cover image features) popularly used by blind steganalysis schemes. The errors induced in the embedded data due to the fact that
the stego signal must be advertised in a specific format such as JPEG,
are dealt with by the use of erasure and error correcting codes. For the
presented JPEG steganograhic scheme, it is shown that the detection
rates of recent blind steganalysis schemes are close to random guessing,
thus confirming the practical applicability of the proposed technique. We
also note that the presented steganography framework, of hiding in randomized locations and using a coding framework to deal with errors, is
quite simple yet very generalizable.
Key words: data hiding, error correcting codes, steganalysis, steganography, supervised learning.
1
Introduction
Secure communication of a secret message has always been important to people,
and it is not surprising that steganography, the art of communicating without
revealing its existence, as well as cryptography, the art of concealing the meaning
of a message, have a rich history. In this paper, we consider the problem of
secure steganography via hiding information in digital images. In steganography,
a message signal is embedded into a host or cover signal to get a composite or
stego signal in such a way that the presence of hidden information cannot be
detected by either statistical or perceptual analysis of the stego signal. In case of
⋆⋆
This research is supported in part by a grant from ONR # N00014-05-1-0816.
2
YASS: Yet Another Steganographic Scheme
active steganography, there is an additional requirement that the hidden data
must be recoverable even after benign or malicious processing of the stego signal
by an adversary.
JPEG is arguably the most popular format for storing, presenting, and exchanging images. It is not surprising that steganography in the JPEG format,
and its converse problem of steganalysis of JPEG images to find ones with hidden data, have received considerable attention from researchers over the past
decade. There are many approaches and software available for JPEG steganography, which include OutGuess [1], StegHide [2], model-based steganography [3],
perturbed quantization [4], F5 [5], and statistical restoration [6, 7].
Approaches for JPEG steganography have focused on hiding data in the
least significant bit (LSB) of the quantized discrete cosine transform (DCT)
coefficients. In order to avoid inducing significant perceptual distortion in the
image, most methods avoid hiding in DCT coefficients whose value is 0. To
detect the presence of data embedded in this manner, steganalysis algorithms
exploit the fact that the DCT coefficient histogram gets modified when hiding
random information bits. Hence recently proposed steganographic approaches
attempt to match, as closely as possible, the original DCT histogram or its
model. Westfield’s F5 algorithm [5] increases, decreases, or keeps unchanged, the
coefficient value based on the data bit to be hidden, so as to better match the
host statistics. Provos’s OutGuess [1] was the first attempt at explicitly matching
the DCT histogram. Sallee proposed a model based approach for steganography
[3] wherein the DCT coefficients were modified to hide data such that they
follow an underlying model. Fridrich et al’s perturbed quantization [4] attempts
to resemble the statistics of a double-compressed image. Statistical restoration
method proposed by Solanki et al [6, 7] can match the DCT histograms exactly,
thus providing provable security so long as only the marginal statistics are used
by the steganalyst.
Many steganalysis schemes (see [8–10]) have been able to successfully detect
the above steganographic techniques that match marginal statistics or models.
They exploit the fact that higher order statistics get modified by data hiding using these stego methods. It is known that, the higher order statistics, in general,
are difficult to match, model, or restore. Recently, blind steganalysis algorithms
[9–16] have been proposed that employ supervised learning to distinguish between the plain cover and stego images, and also identify the particular hiding
algorithm used for steganography. These techniques bank on the fact that there
are some image features that are modified during the embedding process which
can be used as an input to the learning machine. For the success of this approach,
it is crucial that these features are very sensitive to the embedding changes, but
insensitive to the image content. This requires a good model for natural images
against which the suspected stego images can be evaluated.
In spite of the absence of good universal models, recent steganalysis algorithms have been very successful by using a self-calibration method to approximate the statistics of the original cover (see, for example, Pevny and Fridrich
[9, 10], and Dabeer et al [17]). The calibration method typically used for JPEG
YASS: Yet Another Steganographic Scheme
3
steganography is quite simple; a few pixel rows and/or columns are cropped from
the image so as to desynchronize it from the original JPEG grid and the resulting
image is compressed again, which forms a good approximation of the cover image. The results reported in [10], the most recent multi-class JPEG steganalysis
method that employs such self-calibration, are close to perfect: the steganalyst
can determine one out of 6 stego algorithms employed for hiding with a detection
accuracy of more than 95% in most cases, even at low embedding rates.
We present yet another steganographic scheme (YASS), a method for secure,
active, steganography that can successfully resist the aforementioned blind steganalysis schemes. The technique is based on a simple idea of embedding data
in random locations within an image, which makes it difficult for the steganalyst to get a good estimate of the cover image features via the self-calibration
process. This approach works by desynchronizing the steganalyst, which is similar in spirit to what stirmark does for watermarks. Although data is hidden in
randomly chosen blocks in the image, the image must be advertised in JPEG
format. This leads to errors in the recovered data bits, which are dealt with by
using erasure and error correcting codes in a manner similar to [18].
We evaluate the method against several recent blind steganalysis schemes,
which include Farid’s 72-dimensional wavelet-based features [19], Pevny and
Fridrich’s 23-dimensional DCT-based features [9] and 274-dimensional merged
Markov and DCT features [10], a feature vector comprising of histogram of DCT
coefficients, Xuan et al’s statistical moments based spatial domain steganalysis
features [13], and Chen et al’s 324-dimensional JPEG steganalysis feature [20].
We find that the presented method, YASS, is completely undetectable for most
of the embedding configurations, while the competing algorithms, OutGuess and
StegHide, are detectable at the same hiding rates. We also note here that, because we use error correction coding framework, our method provides robustness
against distortion constrained attacks thus enabling active steganography.
The rest of the paper is organized as follows. In Section 2, we motivate the
use of the proposed randomized hiding for steganography. Next, in Section 3,
we introduce and describe our JPEG steganographic technique: YASS. Experimental setup and results are presented in Section 4, followed by the concluding
remarks in Section 5.
2
Resisting Blind Steganalysis
The notion of ǫ-security proposed by Cachin [21] states that a steganographic
scheme is ǫ-secure if the Kullback-Leibler divergence between the cover and the
stego signal distributions is less than a small number ǫ. This definition inherently
assumes that cover signals can be described by “natural” distributions, which
are known to the steganalyst. Statistical steganalysis schemes work by evaluating
a suspected stego signal against an assumed or computed cover distribution or
model. Most recent steganalysis schemes fall in this category [22, 10, 9].
Blind statistical steganalysis schemes use a supervised learning technique on
features derived from plain cover as well as stego signals. This class of meth-
4
YASS: Yet Another Steganographic Scheme
ods has been very successful in detecting steganographic methods available today. For example, detection results presented in [10] and also our own experiments indicate that popular JPEG steganographic schemes such as OutGuess
[1], StegHide [2], model-based steganography [3], and 1D statistical restoration
schemes [6, 7] can be successfully detected. Following are the key ingredients that
contribute to the success of these blind steganalysis schemes.
1. Self-calibration mechanism: Calibration process is used by the blind steganalysis schemes to estimate the statistics of the cover image from the stego
image. For JPEG steganography, this is typically achieved by decompressing
the stego image to the spatial domain followed by cropping the image by a
few pixels on each side and compressing the image again using the same
compression parameters.
2. Features capturing cover memory: Most steganographic schemes hide
data on a per-symbol basis, and typically do not explicitly compensate or
preserve statistical dependencies. Hence, features that capture higher dimensional dependencies in the cover symbols are crucial in detecting the
embedding changes. Cover memory has been shown to be very important to
steganalysis [22], and is incorporated into the feature vector in several ways,
e.g. [23, 15].
3. Powerful machine learning: Use of powerful machine learning techniques
and training with several thousand images ensures that even the slightest
statistical variation in the features is learned by the machine.
The calibration process is perhaps the most important of the above that
allows the steganalyst to get an accurate underlying model of the cover image
despite not having access to it. With this approach, statistical steganalysis is
successful eventhough good universal statistical models for images are not available. In the following section, we discuss a simple approach that can potentially
defeat these steganalysis schemes.
2.1
The Proposed Approach for Steganography
In order to enable secure communication in the presence of blind steganalysis, the
steganographer must embed information into host signals in such a way that no
image features are significantly perturbed during the embedding process. However, we must not forget that the steganalyst must depend on the stego image
to derive the approximate cover image statistics via some sort of self-calibration
process. The steganographer can, instead of (or along with) trying to preserve
the feature vectors, embed data in such a way that it distorts the steganalyst’s
estimate of the cover image statistics. This can practically be achieved using the
following approaches.
1. Hiding with high embedding strength: By embedding data with high
strength, the cover image is distorted so much that the cover image statistics
can no longer be derived reliably from the available stego image. This is
indeed found to be true and reported in recent work by Kharrazi et al [24].
YASS: Yet Another Steganographic Scheme
5
2. Randomized hiding: By randomizing the embedding approach, the algorithm to estimate the cover statistics can be effectively disabled. Things
that can be randomized include the spatial location of hiding, the transform
coefficient to hide, the choice of transform domain, or even the embedding
method. In this manner, the steganalyst cannot make any consistent assumptions about the hiding process even if the embedding algorithm is known to
everyone as per the Kerckhoff’s principle.
There are some obvious disadvantages of using the first approach of hiding
with high strength. First, the likelihood of perceptual distortion is high. Second,
the data can possibly be detected by a steganalyst evaluating the stego image
against a universal image model even if it is not that precise.
The second approach of hiding in a randomized manner is quite appealing, and we explore its simplest realization in this paper: embedding data in
randomized locations within an image. One issue with hiding data in random
locations is the possibility of encountering errors in the hidden bits due to the
fact that the stego image must be shipped or advertised in a standard format
such as JPEG. This is dealt with by the use of erasures and error correction
coding framework previously employed in [18]. In the next section, we describe
yet another steganographic scheme (YASS), a JPEG stegosystem based on the
aforementioned framework.
3
YASS for JPEG Steganography
We now present a JPEG steganography scheme, YASS, that embeds data in 8×8
blocks whose locations are chosen randomly so that they do not coincide with
the 8×8 grid used during JPEG compression. Let the host image be denoted
by a M ×N matrix of pixel values. For simplicity, we assume that the image
is grayscale (single channel); if it is not, we extract its luminance. Below we
describe the main steps involved in this randomized block hiding method.
1. Divide the image into blocks of size B × B, where B, which we call big
block size, is always greater than 8, the size of a JPEG block. Thus we have
N
MB × NB big blocks in the image where MB = ⌊ M
B ⌋ and NB = ⌊ B ⌋.
2. For each block (i, j) (0 ≤ i < MB ,0 ≤ j < NB ), we pseudorandomly select
a 8×8 sub-block in which to hide data. The key for the random number
generator is shared between the encoder and the decoder. The pseudorandom
number generator determines the location of the smaller 8 × 8 block within
the big block. This process is illustrated in Figure 1(a) where four example
blocks are shown, whose top leftmost corner (sx , sy ) is randomly chosen
from the set {0, 1, ..., B − 8}. Figure 1(b) shows the blocks as seen by the
steganalyst who gets out-of-sync from the embedding blocks, and cannot
resynchronize even if the embedding mechanism is known.
3. For every 8 × 8 block thus chosen, we compute its 2D DCT and divide it by
a JPEG quantization matrix at a design quality factor QFh . Data is hidden
in a predetermined band of low frequency AC coefficients using quantization
6
YASS: Yet Another Steganographic Scheme
index modulation. For maintaining perceptual transparency, we do not hide
in coefficients that quantize to zero by the JPEG quantizer (following the
selective embedding in coefficients scheme proposed in [18]).
(a) YASS hiding methodology: Data is hidden in randomly chosen 8×8 blocks within
a big block of size B × B, with B > 8. This
example uses B = 10.
(b) The block structure as seen by a
steganalyst, who gets out-of-sync with
the blocks used during embedding, and
cannot synchronize even if the hiding
method and its parameters are known.
Fig. 1. The embedding method and its detection.
Note that using this approach, we can effectively de-synchronize the steganalyst so that the features computed by him would not directly capture the
modifications done to the image for data hiding (see Figure 1). It should be noted
that with this embedding procedure, we are reducing the embedding rate in two
ways. First, some real estate of the image is wasted by choosing bigger blocks
from which an 8 × 8 block is chosen to hide data. Note that the above framework
can be further generalized to enable lesser wastage, by using larger big blocks
and putting more 8×8 blocks into them. For example, we can use big blocks
of size 33 × 33 and embed in sixteen 8 × 8 blocks within. We report results for
such implementations as well (Section 4). The second cause of decrease in rate
is that since the embedding grid does not coincide with the JPEG grid, there
are errors in the received data which must be corrected by adding redundancy.
This process is briefly described in the following section.
3.1
Coding Framework
In order to deal with the errors caused in the image due to JPEG compression,
we use a coding framework using repeat-accumulate (RA) codes [25], similar
to that proposed in [18]. This framework also allows us to hide in an adaptive
fashion, avoiding coefficients that quantize to zero so as to control the perceptual
distortion to the image.
For every block, we consider an embedding band comprising of first n low
frequency coefficients which forms the candidate embedding band. Data bits are
hidden in a coefficient lying in the band if it does not quantize to zero using the
JPEG quantizer at QFh . Before the hiding process, the bit stream to be hidden
YASS: Yet Another Steganographic Scheme
7
is coded, using a low rate code, assuming that all host coefficients that lie in the
candidate embedding band will actually be employed for hiding. A code symbol
is erased at the encoder if the local adaptive criterion (of being quantized to
zero) for the coefficient is not met. A rate 1/q RA encoder is employed, which
involves q-fold repetition, pseudorandom interleaving and accumulation of the
resultant bit-stream. Decoding is performed iteratively using the sum-product
algorithm [26]. The use of this coding framework for YASS provides the following
advantages.
1. Protection against initial JPEG compression: Use of the coding framework provides error-free recovery of the hidden data after the initial JPEG
compression so that the image can be advertised in the JPEG format.
2. Flexibility in choosing hiding locations: The coding framework allows
us to dynamically select the embedding locations in order to limit the perceptual distortion caused to the host image during hiding. It is well known
that embedding in DCT coefficients that quantize to zero can lead to visible
artifacts in the stego image.
3. Enabling active steganography: The use of error correcting codes also
provides protection against several distortion constrained attacks that an
active warden might perform. The attacks that can be survived include a
second JPEG compression, additive noise, limited amount of filtering, and
so on. This provides a significant advantage over most other stego methods
available in the literature.
4
Experiments and Results
We conduct a comprehensive set of experiments and present the results demonstrating the applicability of the presented approach. First, the results for the
embedding capacity are presented for some standard images (Section 4.1). Next,
in Section 4.2, we present the detection results of our scheme using recent blind
steganalysis methods for a couple of datasets (comprising of several thousand
natural images). We also compare the detection results of our method with those
of OutGuess and StegHide in Section 4.3, and show that our methods clearly
outperform these approaches. As a control experiment, in Section 4.4, we compare our hiding method with a naive standard hiding approach, in which data
is hidden in randomly chosen low-frequency DCT coefficients using standard
JPEG grid, while keeping the same embedding rate.
4.1
Embedding Rates
In Table 1, we list the number of bits that can be hidden in several standard
images using YASS with different embedding parameters. QFh denotes the design quality factor used during hiding, and QFa denotes the output or advertised
image quality factor. Note that for the presented scheme, these two can be different and their values do affect the steganalysis performance (see Section 4.2).
8
YASS: Yet Another Steganographic Scheme
The same notations are used in other tables presented in this paper. Also note
that the number of bits reported in this section are before coding and can be
recovered without any errors, even in the presence of distortion constrained attacks. For all the results presented in this paper, we use 19 low-frequency DCT
coefficients as the candidate embedding band.
Table 1. Number of information bits that can be hidden in some standard images of
size 512×512. The big-block size B = 9.
QF h
50
50
75
60
→ QF a baboon peppers airplane lena couple crowd
→ 50
1453
1295
2128 1702 1655 2979
→ 75
14896
6620
7448 7448 9930 11916
→ 75
1453
1805
2590 2128 2128 3972
→ 75
11916
6620
7448 6620 8512 9930
From this table, we see that the number of bits that can be embedded increases when QF a > QF h , however this also leads to slightly more successful
steganalysis performance (to be shown in Tables 4 and 6), which gives us a
trade-off between the embedding rate and the detection performance. When
QFa equals QFh , there are more physical errors in the channel leading to a
reduced embedding rate.
In Table 2, we study the effect of using different big-block sizes B on the
embedding capacity. Here we also report the bits per non-zero coefficients (bpnc).
It is seen from this table that using lower B provides higher embedding capacity,
which is because we end-up using more real estate of the host image for hiding.
We can ensure even higher percentage of image utilization by using larger bigblocks and putting greater number of 8×8 blocks within. For example, we can
put four 8×8 blocks in a 17×17 block, thus making an effective big-block size
2
of Bef f = 17
2 = 8.5. We experiment with block-size B = (n×8 + 1) and use n
8×8 blocks for hiding, and report the results in Table 3. The embedding rate
improves as we use more image area for hiding, but the detection results get
worse (from the steganographer’s viewpoint, see Tables 4 and 6). Also, using
a lower Bef f does not always guarantee a higher embedding capacity because
of the fact that more errors may be induced due to JPEG compression, which
forces us to use a larger redundancy factor, q, of the RA code.
4.2
Detection Results
The steganographic security of our scheme is evaluated against the following
blind steganalysis schemes. The names in bold are the ones used to denote the
steganalysis schemes in the tables presented in this paper.
1. Farid: 72-dimensional feature vector based on moments in the wavelet domain [19].
2. PF-23: Pevny and Fridrich’s 23-dimensional DCT feature vector [9].
3. PF-274: Pevny and Fridrich’s 274-dimensional feature vector that merges
Markov and DCT features [10].
YASS: Yet Another Steganographic Scheme
9
Table 2. Hiding rates for the 512×512 Lena image for different big-block sizes B.
bpnc denotes bits per non-zero coefficients and databits denotes the number of hidden
information bits.
QF h → QF a
B
50 → 50
databits
bpnc
50 → 75
databits
bpnc
75 → 75
databits
bpnc
60 → 75
databits
bpnc
9
1702
0.072
7448
0.213
2128
0.059
6620
0.187
B
databits
bpnc
databits
bpnc
databits
bpnc
databits
bpnc
10
1497
0.064
5491
0.159
2059
0.057
5491
0.158
B
databits
bpnc
databits
bpnc
databits
bpnc
databits
bpnc
12
882
0.038
4189
0.118
1457
0.040
3724
0.105
B
databits
bpnc
databits
bpnc
databits
bpnc
databits
bpnc
14
464
0.020
2736
0.077
794
0.022
2462
0.069
Table 3. Hiding rates for the 512×512 Lena image for larger big-block sizes B = 9,
25, 49, 65 and 81, which can incorporate several 8×8 blocks.
QF h → QF a
B
50 → 75
databits
bpnc
75 → 75
databits
bpnc
60 → 75
databits
bpnc
9
B
25
B
49
B
65
B
81
7448 databits 7834 databits 8064 databits 7963 databits 8064
0.213 bpnc 0.224 bpnc 0.235 bpnc 0.228 bpnc 0.229
2128 databits 2350 databits 2341 databits 1837 databits 1577
0.059 bpnc 0.065 bpnc 0.065 bpnc 0.051 bpnc 0.044
6620 databits 7050 databits 8064 databits 7166 databits 7258
0.187 bpnc 0.200 bpnc 0.231 bpnc 0.203 bpnc 0.205
4. DCT hist.: Histogram of DCT coefficients from a low-frequency band [7].
5. Xuan-39: Spatial domain steganalysis proposed by Xuan et al [13] (a 39dimensional feature vector).
6. Chen-324: JPEG steganalysis based on statistical moments of wavelet characteristic functions proposed by Chen et al [20].
Since we decompress the images at the time of hiding, it can be argued that
spatial domain steganalysis schemes may be able to detect the presence of embedded data. Hence we include a recent spatial domain steganalysis scheme in
our tests.
We conduct our experiments on two datasets: the first having 4500 images in
compressed (JPEG) format and the other having 2000 images in uncompressed
TIFF format. For each dataset, we use half the data for training and the other
half for testing. The training and testing sets have an equal number of cover
and stego images. The idea behind using datasets in different formats is to
study the effect of using already compressed images verses uncompressed images
for our method since it hides in a desynchronized 8×8 grid. As we see later,
there is not much difference in the observed detection rates for the two datasets.
As in all published blind steganalysis approaches, we train a support vector
machine (SVM) on a set of known stego and cover images and use the threshold
thus obtained, to distinguish between the cover and stego images at the time of
testing.
10
YASS: Yet Another Steganographic Scheme
The SVM classifier has to distinguish between two class of images: cover (class
‘0’) and stego (class ‘1’). Let X0 and X1 denote the events that the actual image
being observed belongs to classes ‘0’ and ‘1’, respectively. On the detection side,
let Y0 and Y1 denote the events that the observed image is classified as belonging
to classes ‘0’ and ‘1’, respectively. We use the probability of detection, Pdetect as
our evaluation criteria, which is defined as follows.
Pdetect = 1 − Perror
Perror = P (X0 )P (Y1 |X0 ) + P (X1 )P (Y0 |X1 )
1
1
1
= PF A + Pmiss , for P (X0 ) = P (X1 ) =
2
2
2
where PF A = P (Y1 |X0 ) and Pmiss = P (Y0 |X1 ) denote the probability of false
alarm and missed detection respectively. Note that the above equation assumes
an equal number of cover and stego images in the dataset. For the steganalysis
results, we report Pdetect upto 2 significant digits after the decimal point. An
uninformed detector can classify all the test images as stego (or cover) and get
an accuracy of 0.5. Thus, Pdetect being close to 0.5 implies nearly undetectable
hiding, and as the detectability improves, Pdetect should increase towards 1.
Table 4. JPEG dataset: Steganalysis results for randomized block based hiding, when
big-block size B is varied. It can be seen that the detection is random for most of the
configurations.
QF h (used
QF a
Steganalysis Detection accuracy: Pdetect
for hiding) (advertised QF) Method
B=9 B=10 B=12 B=14
50
50
Farid
0.52
0.51
0.52
0.51
50
75
Farid
0.55
0.55
0.54
0.51
75
75
Farid
0.52
0.51
0.52
0.51
50
50
PF-23
0.56
0.55
0.54
0.54
50
75
PF-23
0.59
0.59
0.56
0.60
75
75
PF-23
0.53
0.57
0.53
0.52
50
50
PF-274
0.58
0.56
0.53
0.55
50
75
PF-274
0.77
0.79
0.74
0.65
75
75
PF-274
0.59
0.60
0.62
0.54
50
50
DCT-hist
0.53
0.53
0.51
0.53
50
75
DCT-hist
0.64
0.64
0.60
0.54
75
75
DCT-hist
0.55
0.54
0.55
0.53
50
50
Xuan-39
0.54
0.56
0.54
0.51
50
75
Xuan-39
0.63
0.64
0.57
0.53
75
75
Xuan-39
0.52
0.54
0.53
0.52
50
50
Chen-324
0.57
0.51
0.55
0.54
50
75
Chen-324
0.75
0.65
0.60
0.55
75
75
Chen-324
0.54
0.55
0.53
0.53
In Tables 4 and 5, we present the detection accuracy obtained on the JPEG
and TIFF image dataset respectively. Note that even for the TIFF dataset, the
YASS: Yet Another Steganographic Scheme
11
Table 5. TIFF dataset: Steganalysis results for randomized block based hiding, when
the big-block size B is varied. It can be seen that the detection is random for most of
the configurations.
QF h (used
QF a
Steganalysis Detection accuracy: Pdetect
for hiding) (advertised QF) Method
B=9 B=10 B=12 B=14
50
50
Farid
0.51
0.51
0.51
0.51
50
75
Farid
0.53
0.51
0.52
0.51
75
75
Farid
0.50
0.51
0.51
0.51
50
50
PF-23
0.53
0.55
0.53
0.53
50
75
PF-23
0.59
0.66
0.53
0.53
75
75
PF-23
0.54
0.51
0.53
0.53
50
50
PF-274
0.52
0.56
0.55
0.51
50
75
PF-274
0.72
0.81
0.65
0.60
75
75
PF-274
0.56
0.56
0.52
0.53
50
50
DCT-hist
0.51
0.53
0.52
0.48
50
75
DCT-hist
0.60
0.59
0.56
0.56
75
75
DCT-hist
0.52
0.52
0.51
0.52
50
50
Xuan-39
0.52
0.51
0.52
0.48
50
75
Xuan-39
0.57
0.60
0.53
0.48
75
75
Xuan-39
0.47
0.51
0.49
0.47
50
50
Chen-324
0.55
0.55
0.53
0.53
50
75
Chen-324
0.65
0.64
0.56
0.54
75
75
Chen-324
0.55
0.51
0.51
0.50
output is always a JPEG image at an advertised quality factor. For this dataset,
the plain cover images are JPEG compressed at QF a during the training as well
as testing. It can be seen from the tables that our scheme is undetectable using
any of the steganalysis features. The only time the detection is not completely
random is when the design quality factor is lower than the advertised one (QFh =
50 and QFa = 75). We also present, in Table 6, the steganalysis results when
larger big-block sizes are used, so as to incorporate more 8×8 blocks within.
Pdetect remains close to 0.5 in most cases, but for PF-274, it increases as B
increases (in Table 6), since more area of the image is used for hiding when
employing larger big-blocks. Note that since we are using a set of images and
the embedding rate varies for individual images, we cannot provide bpnc values
in these tables.
4.3
Comparison with Competing Methods
In Table 7, we present a comparison of our steganographic scheme, YASS, to
OutGuess[1] and StegHide[2]. To enable fair comparison, we must use the same
hiding rate for all the schemes. This is complicated by the fact that YASS uses an
error correcting code whose redundancy factor, q, determines the actual number
of information bits hidden in the image. Experiments indicate that, for images in
our dataset, the redundancy factor required to ensure zero BER was in the range
10-40, which depends on the particular image and the level of JPEG compression
12
YASS: Yet Another Steganographic Scheme
Table 6. Using larger big-block size B: Steganalysis results for randomized block based
hiding on the TIFF image dataset, for block-size B = 9, 25 and 49. For 9×9, 25×25
and 49×49 blocks, we use 1, 9 and 36 8×8 sub-blocks respectively for hiding.
QF h (used
QF a
Steganalysis Detection accuracy: Pdetect
for hiding) (advertised QF) Method
B=9
B=25
B=49
50
50
Farid
0.51
0.50
0.50
50
75
Farid
0.53
0.51
0.52
75
75
Farid
0.50
0.49
0.51
50
50
PF-23
0.53
0.53
0.57
50
75
PF-23
0.59
0.58
0.67
75
75
PF-23
0.54
0.53
0.53
50
50
PF-274
0.52
0.57
0.59
50
75
PF-274
0.72
0.73
0.78
75
75
PF-274
0.56
0.58
0.68
50
50
DCT-hist
0.51
0.50
0.51
50
75
DCT-hist
0.60
0.56
0.50
75
75
DCT-hist
0.52
0.51
0.50
50
50
Xuan-39
0.52
0.52
0.52
50
75
Xuan-39
0.57
0.53
0.50
75
75
Xuan-39
0.47
0.52
0.51
50
50
Chen-324
0.55
0.53
0.52
50
75
Chen-324
0.65
0.58
0.57
75
75
Chen-324
0.55
0.52
0.51
involved. Hence we present results where the amount of data embedded using
OutGuess and StegHide corresponds to the hiding rate obtained using q = 10
and q = 40. It can be seen from the table that both OutGuess and StegHide
are almost completely detectable (especially when using PF-23, PF-274, and
DCT-hist features), but YASS is not detectable at equivalent hiding rates.
We also experimented with the F5 steganographic scheme [5], which uses
matrix embedding (see for example [27]), and found that this scheme is also
undetectable at equivalent embedding rates using PF-274 and PF-23 features.
Matrix embedding allows hiding at embedding efficiencies (defined as number
of bits hidden for every change made to the host symbols), potentially much
higher than the trivial efficiency of 2 bits per change (bpc). Thus for passive
steganography, schemes that employ matrix embedding (such as F5) can enable
undetectable hiding at embedding rates equivalent to YASS. However, an added
advantage of YASS is that it can provide protection against active adversaries
by its use of error correcting codes.
4.4
Comparison with Standard Hiding at Same Rate
In the previous sections, it is seen that our proposed steganographic approach
performs quite well against blind steganalysis schemes that we have tested against.
We must, however, note that the improved steganographic security comes at the
YASS: Yet Another Steganographic Scheme
13
Table 7. Steganalysis results for comparing the randomized block based scheme
1
1
and 40
, for the
(YASS) with OutGuess and Steghide schemes, used at rates of 10
TIFF image dataset. For OutGuess and Steghide, the images are JPEG compressed
using a quality factor of QF a before being presented to the steganographic scheme.
Note that the QF h parameter is applicable only for the YASS scheme.
QF h QF a Steganalysis
Method
50 50
Farid
50 75
Farid
75 75
Farid
50 50
PF-23
50 75
PF-23
75 75
PF-23
50 50
PF-274
50 75
PF-274
75 75
PF-274
50 50 DCT-hist
50 75 DCT-hist
75 75 DCT-hist
Detection accuracy: Pdetect
1
1
1
1
YASS OutGuess- 10
Steghide- 10
OutGuess- 40
Steghide- 40
0.51
0.74
0.50
0.77
0.52
0.53
0.59
0.50
0.50
0.55
0.50
0.59
0.50
0.50
0.55
0.53
0.98
0.78
0.97
0.80
0.59
1.00
0.99
0.99
0.99
0.54
1.00
0.99
0.99
0.99
0.52
1.00
0.98
1.00
0.96
0.72
1.00
1.00
1.00
1.00
0.56
1.00
1.00
1.00
1.00
0.51
0.95
0.59
0.94
0.60
0.60
1.00
0.91
1.00
0.93
0.52
1.00
0.91
1.00
0.93
cost of reduced embedding rate. To further investigate whether the good performance of YASS is simply because of reduced rate or not, we present another
simpler extension of the idea of embedding in random locations: we now embed
data in randomly chosen low-frequency AC DCT coefficients computed using
the original JPEG grid. This approach would incur minimal distortion to the
original cover during the hiding process, and hence would be an ideal control
experiment for testing our scheme. The results are reported in Table 8, where
we compare the YASS embedding scheme with this randomized frequency (RF)
scheme for three hiding rates: 2 out of 19, 1 out of 19, and 1 out of 38 coefficients. It can be seen that the naive RF scheme performs quite well, however,
the performance of YASS is consistently better.
We end this section by noting that the reported results (for RF) are for
embedding at trivial embedding efficiency of 2 bpc. As stated earlier, the detectability can be reduced by causing fewer changes to the DCT coefficients by
using matrix embedding. One of the key factors that determines detectability is
the change rate. This change rate can be thought of as the encoders’ “budget”,
which can be “used” either way: to provide robustness by using redundancy and
giving up some embedding efficiency, or to improve the embedding efficiency
via matrix embedding but causing an increase in the fragility of the system.
YASS goes the first way: it can provide robustness against an active adversary
by using error correcting codes but the embedding efficiency is quite low. However, we emphasize that for passive warden steganography, an equivalent level
of undetectability can be achieved at the same embedding rates using matrix
embedding.
14
YASS: Yet Another Steganographic Scheme
Table 8. Comparison of steganalysis results for randomized block (RB) hiding (i.e.,
YASS) with B = 9, and randomized frequency (RF) based hiding.
QF h (used
QF a
Steganalysis
Detection accuracy: Pdetect
for hiding) (advertised QF) Method
RB (YASS) RF 1/38 RF 1/19 RF 2/19
50
50
Farid
0.51
0.66
0.65
0.67
75
75
Farid
0.50
0.52
0.58
0.67
50
50
PF-23
0.53
0.69
0.83
0.92
75
75
PF-23
0.54
0.58
0.72
0.83
50
50
PF-274
0.52
0.71
0.79
0.90
75
75
PF-274
0.56
0.62
0.75
0.82
50
50
DCT-hist
0.51
0.55
0.68
0.86
75
75
DCT-hist
0.52
0.54
0.59
0.80
5
Conclusions
In this paper, we have demonstrated a simple yet very effective steganographic
approach that provides security against recent blind steganalysis schemes. The
method embeds data in randomly chosen host blocks, thus relying on confusing
the steganalyst’s estimate of the cover statistics, rather than preserving the
host image features. We note that the improved security comes at the cost of
embedding capacity, and in the future, we will investigate schemes that can
significantly increase the hiding capacity while maintaining the steganographic
security. We will also explore other mechanisms for randomizing the embedding
process, such as using randomly chosen transform domains.
References
1. Provos, N.: Defending against statistical steganalysis. In: 10th USENIX Security
Symposium, Washington DC, USA (2001)
2. Hetzl, S., Mutzel, P.: A graph theoretic approach to steganography. In: 9th IFIP
TC-6 TC-11 International Conference, Communications and Multimedia Security.
Volume 3677., Salzburg, Austria (2005) 119–128
3. Sallee, P.: Model-based steganography. In: IWDW 2003, LNCS 2939. (2003) 154–
167
4. Fridrich, J., Goljan, M., Lisonˇek, P., Soukal, D.: Writing on wet paper. In: ACM
Workshop on Multimedia and security, Magdeburg, Germany (2004)
5. Westfeld, A.: High capacity despite better steganalysis (F5 - a steganographic
algorithm). In: Lecture notes in computer science: 4th International Workshop on
Information Hiding. Volume 2137. (2001) 289–302
6. Solanki, K., Sullivan, K., Madhow, U., Manjunath, B.S., Chandrasekaran, S.: Statistical restoration for robust and secure steganography. In: Proc. ICIP. (2005)
II–1118–21
7. Solanki, K., Sullivan, K., Madhow, U., Manjunath, B.S., Chandrasekaran, S.: Probably secure steganography: Achieving zero K-L divergence using statistical restoration. In: Proc. ICIP. (2006) 125–128
8. Wang, Y., Moulin, P.: Steganalysis of block-DCT image steganography. In: IEEE
workshop on Statistical Signal Processing, St Louis, MO, USA (2003)
YASS: Yet Another Steganographic Scheme
15
9. Pevny, T., Fridrich, J.: Multi-class blind steganalysis for JPEG images. In: Proc.
of SPIE, San Jose, CA (2006)
10. Pevny, T., Fridrich, J.: Merging Markov and DCT features for multi-class JPEG
steganalysis. In: Proc. of SPIE, San Jose, CA (2007)
11. Avcibas, I., Sankur, B., Memon, N.: Image steganalysis with binary similarity
measures. In: Proc. ICIP. (2002) 645–648
12. Lyu, S., Farid, H.: Detecting hidden messages using higher-order statistics and
support vector machines. In: Lecture notes in computer science: 5th International
Workshop on Information Hiding. Volume 2578. (2002)
13. Xuan, G., Shi, Y.Q., Gao, J., Zou, D., Yang, C., Yang, C., Zhang, Z., Chai, P.,
Chen, C., Chen, W.: Steganalysis based on multiple features formed by statistical
moments of wavelet characteristic functions. In: Lecture notes in computer science:
7th International Workshop on Information Hiding. (2005)
14. Harmsen, J.J., Pearlman, W.A.: Steganalysis of additive noise modelable information hiding. In: Proc. of SPIE. (2003) 131–142
15. Shi, Y.Q., Chen, C., Chen, W.: A Markov process based approach to effective
attacking JPEG steganography. In: Lecture notes in computer science: 8th International Workshop on Information Hiding. (2006)
16. Wang, Y., Moulin, P.: Optimized feature extraction for learning-based image steganalysis. IEEE Transactions on Information Forensics and Security 2(1) (2007)
31–45
17. Dabeer, O., Sullivan, K., Madhow, U., Chandrasekaran, S., Manjunath, B.: Detection of hiding in the least significant bit. IEEE Transactions on Signal Processing,
Supplement on Secure Media I 52(10) (2004) 3046–3058
18. Solanki, K., Jacobsen, N., Madhow, U., Manjunath, B.S., Chandrasekaran, S.:
Robust image-adaptive data hiding based on erasure and error correction. IEEE
Trans. on Image Processing 13(12) (2004) 1627 –1639
19. Farid, H.: http://www.cs.dartmouth.edu/farid/research/steg.m. (Code for generating wavelet-based feature vectors for steganalysis.)
20. Chen, C., Shi, Y.Q., Chen, W., Xuan, G.: Statistical moments based universal
steganalysis using JPEG-2D array and 2-D characteristic function. In: Proc. ICIP,
Atlanta, GA, USA (2006) 105–108
21. Cachin, C.: An information theoretic model for steganography. LNCS: 2nd Int’l
Workshop on Info. Hiding 1525 (1998) 306–318
22. Sullivan, K., Madhow, U., Chandrasekaran, S., Manjunath, B.: Steganalysis for
Markov cover data with applications to images. IEEE Transactions on Information
Forensics and Security 1(2) (2006) 275–287
23. Fu, D., Shi, Y.Q., Zou, D., Xuan, G.: JPEG steganalysis using empirical transition
matrix in block dct domain. In: International Workshop on Multimedia Signal
Processing, Victoria, BC, Canada (2006)
24. Kharrazi, M., Sencar, H.T., Memon, N.: Cover selection for steganographic embedding. In: Proc. ICIP. (2006) 117–120
25. Divsalar, D., Jin, H., McEliece, R.J.: Coding theorems for turbo-like codes. In:
36th Allerton Conf. on Communications, Control, and Computing. (1998) 201–210
26. Kschischang, F.R., Frey, B.J., Loeliger, H.A.: Factor graphs and the sum-product
algorithm. IEEE Trans. on Info. Theory 47(2) (2001) 498–519
27. Fridrich, J., Soukal, D.: Matrix embedding for large payloads. In: Proc. of SPIE.
(2006) 727–738